OpenClaw 5.22: 4100x Gateway Speedup, No More Slow AI

At 1 a.m., the OpenClaw team dropped version 2026.5.22. The star of this release isn't a new model or a fresh skin—it's a complete teardown-and-rebuild of the Gateway, the heart of the whole system. In one sentence: launching OpenClaw used to feel like waiting for a bus; now it's like sitting at a red light that turns green in seconds.

Even more interesting, this update hides a bunch of "invisible muscle": from faster startup to slimmer Subagent context, to pulling Discord voice calls in as a data source—all hard work obsessed with everyday experience. Let's go through it one by one and see what you should be using today.

Just how much faster is the Gateway this time?

Longtime users know the drill: the Gateway is OpenClaw's main gate—every Agent call, model route, Channel message, and Skill load has to pass through it. The old problem: every startup, every config refresh, every new plugin meant scanning the entire disk and re-reading all Channels, Plugins, and Manifests—enough disk IO to make your drive audibly weep.

5.22 replaces all of that with "snapshot-style caching":

Plugin metadata: instead of re-statting every file on each launch, it now reuses a process-level immutable snapshot—startup reads of config/models/Channels go straight to memory
Plugin SDK aliases: caches the alias map and skips repeated file traversal
Startup idle tasks: lazy-loaded on demand; Gateway health/ready no longer waits on empty handler trees
ACPX probe: now lazy—it doesn't run unless it's actually used
macOS Linuxbrew PATH: skips irrelevant probes, saving a few hundred milliseconds of stat calls at startup

In official tests, on hot paths like the Channel catalog that "repeatedly read the same config," single calls dropped from milliseconds straight down to microseconds,with some scenarios hitting a 4100x speedup. This isn't clickbait—it's a physics-level win from swapping disk IO for memory hits.

Meeting Notes: Turning meetings into Agent input

There's also a deeply hidden easter egg in this release—theMeeting Notes plugin.

It's a standalone source plugin, not bundled in the main package, so you install it separately. It does one thing: turn the act of "having a meeting" into structured data to feed your Agent.

It supports two modes:

Automatic capture: starts recording and transcribing automatically when a meeting begins
Manual import: pasting in a transcript works too

The first natively supported live source is Discord voice. In other words, start a voice meeting on Discord and OpenClaw turns it into notes in real time—your Agent can answer questions, generate minutes, and follow up on action items based on what was actually said.

From the CLI you can query it like this:

openclaw meeting-notes list
openclaw meeting-notes get <id>

Note that access is read-only; writes are managed by the plugin itself. It's a restrained design—you can't mess with the raw data, but your Agent can always consume it.

Subagents finally learned to talk less

Veterans know the pain: dispatching a Subagent used to stuff the main Agent's SOUL.md, IDENTITY.md, USER.md, MEMORY.md, HEARTBEAT.md, and more into its context. The result: context bloat, plus sub Agents that kept drifting off-script with lines like "I'm just a workhorse."

5.22 changes the default:Subagent bootstrap now only carries AGENTS.md and TOOLS.md.

Persona, identity, user profile, memory, heartbeat, init files—none of it gets shipped by default anymore. Sub Agents get pure working context and just do the job.

That means:

Lower token usage (my own testing saved 30%–50% on the same task)
No more character bleed (each Subagent stops announcing it's a "workhorse")
Cleaner multi-Agent collaboration

If you genuinely need some of the main Agent's context, you now have to pass it explicitly—this is both a safety upgrade and a performance upgrade.

Media understanding: Gemini CLI stops guessing

5.22 reworked the media-recognition fallback chain.

Previously, when OpenClaw saw an image or a video, it would first probe whether Gemini CLI could run. Not anymore—it goes straight to the official provider API you configured, with Antigravity CLI demoted to "last-resort fallback."

In practice:

You've configured an image/video understanding API with OpenAI / Anthropic / Zhipu / MiniMax → it gets used directly
None configured → only then does Antigravity CLI step in as the fallback

This avoids a common pitfall: you've set up a primary provider, but something inexplicably routes to Gemini CLI and throws a permission error.

Package slimming: npm tarball drops the docs images

@openclaw/openclaw This time the npm package has kicked all documentation images and static assets out of the tarball.

For end users:

npm install is faster
node_modules is smaller (some users report tens of MB saved)
No impact on runtime doc search or CLI behavior

More and more of these fine-grained optimizations show the team is taking engineering quality seriously.

A pile of documentation patches

The docs updates this time went off like firecrackers. Here are the highlights:

WhatsApp QR / 408 recovery: salvation for anyone whose QR scan kept failing
Cron output language hint: scheduled tasks can now specify a language
Gateway upstream 403 troubleshooting: don't panic if the reverse proxy is suspicious of you
GLM unified to Z.AI naming: Zhipu's models no longer confuse anyone
Upstash Box install guide: a new deployment target
Codex migration prerequisites: a complete guide for switching over from Codex
Bitwarden SecretRef setup: safer secret management
EasyRunner deployment docs: how to spin up a Gateway in one shot
Chinese memory management guide: great news for native speakers

Especially worth mentioning:The MEDIA directive specis now properly documented. Going forward, sending images and audio with the MEDIA: format will behave much more predictably across Channels.

Practical advice for the working crowd

If you're upgrading today, keep these in mind:

Upgrade command:

npm install -g @openclaw/openclaw@latest
openclaw gateway restart

Verify startup got faster:

time openclaw gateway status

Compare startup times before and after the upgrade—the difference is visible to the naked eye.

Check Subagent behavior: if you relied on Subagents automatically pulling in MEMORY.md or SOUL.md, pass it explicitly at call time, or your sub Agent will lose its memory.

Meeting Notes install:

openclaw plugin add meeting-notes

Then configure the Discord voice source and you're ready to go.

Why this update matters

OpenClaw's update cadence over the past month or two has been relentless, but the core signal in 5.22 is:the team has started grinding on underlying performance and experience detailsinstead of just piling on new features.

A 4100x Gateway speedup sounds absurd, but it targets the most painful daily gripe: slow, janky startup. The Subagent context diet fixes "AI character bleed and burning money on tokens." Meeting Notes is a new attempt at pushing AI Agents from "passive responders" to "active consumers of real-world data."

Each change looks small on its own; together, they're a leap in experience.

If you're still on the fence about switching to OpenClaw, 5.22 is a great version to jump in with. If you're already using it, upgrade—there's no reason not to.